August 29th 2026
Curated AI news and stories.
OpenAI is cutting Cursor off from its models over the SpaceX deal
OpenAI told SpaceX on Thursday it will end the contract that supplies its models to Cursor, the AI coding tool SpaceX is buying for $60 billion, with a proposed shutoff of November 12th. The reason it gave is not about Cursor's product but about who now owns it: OpenAI says it "cannot be confident that SpaceX will use our technology within our terms of service," and cited Twitter breaking its OpenAI contract after Elon Musk's takeover and Musk's own sworn testimony this year admitting xAI had violated those terms. Cursor was one of OpenAI's earliest customers, took seed money from OpenAI's Startup Fund in 2023, and OpenAI invested in its $29.3 billion . CEO Michael Truell said OpenAI models are roughly 5% of Cursor's traffic and that the company is trying to resolve it, then asked the question the cutoff actually raises: whether a lab that sells a competing coding can be "neutral infrastructure" for the tools built on top of it. Every developer who wired an OpenAI model into a Cursor workflow now has until November to find out what happens when the supplier and the rival are the same company. SourcesAB
Trump's next chip tariff would hit the servers, and scrap the data-center exemption
The Trump administration is weighing a second round of semiconductor tariffs that would tax not just chips but the products built with them: laptops, gaming consoles, and the data-center servers the AI buildout runs on. The detail that matters is what Commerce officials have signaled in private: the carve-outs attached to January's 25% tariff, which spared data centers, R&D, startups and consumer devices, may not carry over. Commerce Secretary Howard Lutnick favors capping duty-free chip imports at a volume tied to each company's committed US production, which turns the tariff into a lever on where the hardware gets made. The Computer and Communications Industry Association told Politico the change could jeopardize the buildout by raising costs into contracts already signed, and compared the scale to the transcontinental railroad. No final schedule exists yet, and the framework could still shift. What is priced today is uncertainty over the input cost of every rack ordered this year. SourcesBB
Andreessen Horowitz raised $1.1 billion to build the hardware AI runs on
Andreessen Horowitz closed a $1.1 billion Machine Age fund, its first vehicle aimed entirely at the physical layer of AI: processors, memory, networking, storage, robotics, data centers, and what general partner Raghu Raghuram called "full systems for running AI, from data centers to robotics to home AI appliances." The pitch is that the bottleneck has moved. "Every time we have one of these technical epochs, it puts pressure on the infrastructure, but none of us have ever seen it this dramatic," said Martin Casado, who will lead the strategy with Raghuram across early and growth stages. The firm that made its name arguing software eats the world just raised a ten-figure fund on the premise that the constraint is now everything software cannot conjure: capacity, , and power booked years out. It is the clearest signal yet that the smart money thinks the returns for the next few years sit in the parts of the stack you can drop on your foot. SourcesAB
CXMT sued the Pentagon, a week after a court told the Pentagon it lost one
ChangXin Memory Technologies, China's largest maker, sued the US Department of Defense on Friday over its designation as a "Chinese military company," calling the listing arbitrary, unsupported by evidence, and a violation of its due-process rights. CXMT says it designs and sells its chips for civilian and commercial use, and that it spent more than a year giving the Pentagon information to contest the label, at one point receiving a February notice that it would be removed, only for the department to withdraw that notice the same day. The suit, filed in the DC District Court, follows Alibaba's June challenge to the same list, and lands eight days after a federal judge ruled the Pentagon's blacklisting of Anthropic illegal. The pattern is now visible from both sides of the Pacific: the designation power the department treated as discretionary is being met in court by companies with the revenue to outlast an appeal, and CXMT, whose is about to ship in a flagship phone, is one of them. SourcesBB
Nine in ten executives say three years of AI changed nothing at their firm
A survey of nearly 6,000 senior executives across the US, UK, Germany and Australia found that more than 90% report no effect of AI on employment at their own company over the past three years, and 89% report no effect on labor productivity. This is not a survey of skeptics: 69% of the firms use AI, adoption runs to 78% in the US, and two-thirds of the executives use it themselves, for about 1.5 hours a week. The gap is between what has landed and what is expected: the same executives predict AI will raise their productivity 1.4% and cut their employment 0.7% over the next three years, while their employees expect headcount to grow. The paper, "Firm Data on AI" from the National Bureau of Economic Research, is the cleanest read yet on the announcement-versus-deployment gap, and the number to hold onto is the one about the past. The layoffs attributed to AI, the buildout financed against it, and the fund raised yesterday to pour concrete for it are all bets on a productivity gain that, by the accounts of the people who run these companies, has not yet arrived. The editorial takes this up. SourcesAB
SK Hynix broke ground in Indiana, and made-in-America HBM starts in 2029
SK Hynix held a groundbreaking Thursday for a more than $4 billion advanced-packaging plant in West Lafayette, Indiana, its first HBM production base in the United States, with mass production of next-generation memory scheduled for the second half of 2029. The facility packages rather than fabricates: DRAM wafers made in South Korea will be shipped over to be stacked, bonded and tested into finished HBM for American customers, a step that puts the highest-value part of the memory , and its tariff exposure, on US soil. The plant is designed for about 1,000 workers at full capacity. SK Hynix holds 58% of the HBM market by revenue, ahead of Samsung and Micron at 21% each, which is why the location decision is a strategic one and not a jobs announcement: the company that supplies most of the memory in every AI accelerator is deciding it needs to stand physically inside the market it sells to. The 2029 start date is also the tell on how tight this stays. SourcesAB
A Unitree humanoid can be rooted over Bluetooth, and the exploit spreads on its own
Security researcher Olivier Laflamme disclosed two vulnerabilities in Unitree's G1 robot that chain into unauthenticated root access from within Bluetooth range, and the exploit, named UniBLEed, is wormable: once one G1 is compromised, it can spread to the next G1 nearby, and so on. One flaw let a nearby device pull the robot's unique encryption key without pairing, because Unitree's cloud never checked whether the requester owned the serial number; the other used a path-traversal bug to write a file that then executes as root. The two CVEs cover a machine that is sold to universities, factories and researchers and increasingly stands in the same rooms as people. A humanoid is a computer with actuators and a hundred kilograms of reach, and the security model here is worse than a consumer router's. This is the second robot-security disclosure this week to describe a physical machine an attacker can take over remotely, and the industry racing to put these bodies on factory floors has not yet treated their firmware as a safety system. SourcesBA
Jio Platforms cleared its IPO, and India's biggest listing ever is now live
India's markets regulator gave Jio Platforms, the digital arm of Mukesh Ambani's Reliance Industries and the home of its AI ambitions, final clearance to launch an initial public offering that could raise about $3.8 billion and value the company near $137 billion. It would be the largest IPO in India's history, surpassing Hyundai Motor India's 2024 listing, and it carries the backing of both Meta and Google, which took stakes in the business during its telecom rise. The proceeds go mostly to repaying Jio's own borrowings. The listing matters beyond its size because Jio is the distribution layer for AI to the next several hundred million users: a telecom and digital-services platform with the reach to put models in front of a market OpenAI and Meta are both staffing up to reach. A company this large going public also tells you the Asian listing window, which Moonshot and others are eyeing, is open. SourcesBB
Xiaomi will put China's first flagship-grade homegrown memory in a phone
Xiaomi confirmed its 18 Fold, due in September, will ship with LPDDR6 memory from CXMT, the first time a Chinese-made memory chip has gone into a flagship-class smartphone processor. The phone pairs it with Xiaomi's own 3-nanometer Xring O3, built on TSMC's line, and the O3 is billed as the first mobile SoC to support LPDDR6, at up to 113.8GB/s of bandwidth. The significance is not the phone; it is that CXMT, the company that sued the Pentagon on the same day this shipped, has reached mass production of a memory part good enough for a top-tier device. China's memory independence has been a projection for years, dependent on a domestic maker crossing from commodity DRAM into leading-edge parts. This is the crossing, in a consumer product, at exactly the moment the US is deciding whether to tax the servers that memory eventually goes into. SourcesBB
GLM-5.3's weights are open, but the license is aimed at the hyperscalers
Z.ai's flagship GLM-5.3 shipped its on Friday, and the license came with them tells you what the company is actually worried about. Instead of the permissive MIT terms it used for the earlier GLM-5.2, Z.ai wrote a new GLM-5.3 license that requires any company with more than $10 billion in revenue over any 12 months to pass the company's own security review before hosting the model commercially. Everyone below that line runs it freely. The gate exists because GLM-5.3 scores 84.5% on CyberGym, a vulnerability-finding , a mark Z.ai says its pushed higher than it expected, and the model is already reported to have found a serious flaw in Cursor. So the strongest open-weights coding model in the world is free to a startup and conditional for a cloud giant, which inverts the usual shape of a software license and reads as a small Chinese lab writing terms for Amazon and Microsoft rather than the other way around. SourcesBB
Bill Gates wants a robot tax and jobs the AI cannot take
Bill Gates published a long essay arguing that AI will eliminate many entry- and mid-level jobs and proposing two blunt tools to slow it. The first is a "robot tax": today an employer who hires a person pays payroll tax, while one who buys a machine writes it off as an expense, and Gates would tax the automation to fund retraining and a stronger safety net. The second is a category he calls "Human Reserved," borrowed from the idea of protected nature reserves: roles society deliberately keeps for people because the loss would be too great, or because the work should come from a human, like delivering a terminal diagnosis. He floats education and mental health care as candidates and, in the most aggressive version, puts the ceiling near 40% of jobs. Gates is careful that he does not know who would enforce any of it. The proposals land the same week the NBER survey says the displacement has not shown up yet, which is the argument: the time to design the floor is before the fall, not after. SourcesBB
Alibaba opened its first South American data centers, and brought the AI stack
Alibaba Cloud switched on its first cloud region in Brazil on Friday, two data centers with a suite of agentic AI services, its debut in South America and the latest node in a network now spanning 106 availability zones across 31 regions. The expansion is part of a $53 billion global AI-infrastructure push, and Brazil is a deliberate choice: a large market with data-localization rules that a domestic region satisfies, and one where American clouds are the incumbents to displace. The move also carries the tension every Chinese cloud abroad now does, between Brazil's own data-protection law and China's intelligence statutes, and Brazilian enterprises weighing the region will have to price that. For a week that opened with CXMT suing the Pentagon and a US tariff threat, the counter-move is visible here: while Washington debates how to wall the hardware in, Alibaba is laying the software layer down in the markets the export fight does not reach. SourcesAB
Australia says data centers must run on renewables, and the states are fighting it
Australia's federal government moved to require new data centers to fully offset their electricity with renewable energy, with Climate and Energy Minister Chris Bowen saying there would be "no exceptions and no carve-outs," and centers made to buy renewable certificates and prove they can flex demand off the grid at peaks. The policy is a direct answer to a projection that data-center electricity demand will rise roughly sevenfold over the coming decade. It is also already contested: Queensland and the Northern Territory pushed back at a National Cabinet meeting on August 26th, and reporting suggests the enabling legislation may not carry a hard renewable mandate after that fight, with the rule changes headed to the energy ministers' council in September. The template being argued over is the one every grid will face, because the buildout's real constraint is power, and the question of who supplies it, at what carbon cost, and who pays when it strains the grid is now a live political fight on at least three continents. SourcesBB
Construction unions are turning on the politicians who oppose data centers
The backlash against data centers has opened a fault line inside organized labor, and the building trades have picked their side. Construction unions are pushing back on local opposition and, in some cases, threatening to withhold campaign support from politicians who oppose data-center projects, because the buildout means thousands of jobs for electricians, pipefitters and operating engineers, The Wall Street Journal reports. The split runs through the AFL-CIO: President Liz Shuler has to hold together members who see the centers as work and members who fear the technology those centers train will come for their own jobs next. The unions also supply something developers need as much as labor: political cover at the city-council meetings where an unpopular project lives or dies. The people who pour the foundations for AI have become its most organized local constituency, and they are redrawing which candidates are safe to oppose it. It is a concrete example of the buildout creating real jobs now, whatever the models do to jobs later. SourcesBB
A billionaire's WSJ op-ed was written by AI, and the paper decided that is fine
Stanley Druckenmiller confirmed that his Wall Street Journal opinion essay attacking Treasury Secretary Scott Bessent's bond strategy was written using AI, after the detection tool Pangram flagged it. "Of course I used AI," he said, comparing it to using a calculator. The Journal's editorial-page editor Paul Gigot backed him, saying the test is whether a piece reflects the author's genuine argument and whether the author has the standing to make it, and that the paper will not require contributors to disclose AI use. It is a live inflection for how newsrooms handle this: days earlier the Financial Times corrected an undisclosed AI-condensed column, and the New York Times requires human-made freelance work. The Journal has now drawn the opposite line, that a named contributor with credibility can publish AI prose as his own opinion without saying so. For readers, the byline no longer tells you whose sentences you are reading, only whose argument it is claimed to be. SourcesBB
Texas pulled state money from Flock's AI cameras before a report could run
Governor Greg Abbott ordered all Texas state agencies to stop funding Flock's AI-powered license-plate cameras on Thursday night, hours before The Texas Tribune published an investigation showing a state agency had put at least $30 million into building out the surveillance network through grants to local police. The Motor Vehicle Crime Prevention Authority has helped install at least 3,200 Flock cameras since 2023, including nearly 1,200 by the Department of Public Safety under a $15.9 million contract. Abbott's office noted most funding for the cameras is federal, which the order does not touch. The move follows a wave of Texas cities canceling their own Flock contracts over data-sharing and privacy. It is a small, concrete instance of the surveillance-AI backlash reaching a red-state governor, and it lands the same day Abbott's state is fighting over data-center approvals: the physical-AI buildout and the watching-AI buildout are both meeting local limits at once. SourcesBB
OpenAI's own study says the tool helps, but the thinking has to be taught
OpenAI published a randomized trial, run with Bocconi University, that split more than 1,000 first-year students four ways: ChatGPT access, training in causal reasoning, both, or neither. Access to the model raised the quality and coherence of the work, lifting rubric scores by nearly a point. The reasoning training did something the model did not: it produced more original ideas, the kind of variety the polished ChatGPT drafts missed. Students who got both kept the structure and gained the originality. The finding is unusually honest for a company-sponsored study, because it says the tool alone smooths the prose without deepening the thought, and that the deepening comes from teaching people to reason, not from handing them a better model. For every school deciding whether to ban ChatGPT or issue it, the trial points at a third answer: the gain is real but conditional on the human skill the tool is most often accused of replacing. SourcesAB
Cerebras will stack DRAM on the wafer, chasing the same memory everyone else needs
At Hot Chips, Cerebras laid out a roadmap in which its next-generation CS-6 stacks DRAM directly on top of the compute-and-SRAM wafer for the first time, using 3D packaging to hold more of a model on each system without giving up the on-wafer locality that makes its fast. It also detailed Nexus, a rack-scale architecture it says triples performance over the prior generation. The strategic read is the same shortage that runs through the rest of the week: even the company whose whole pitch is that it does not need conventional memory hierarchies is now reaching for stacked DRAM to expand capacity, because there is no way to serve today's models without more memory close to the compute. Wafer-scale was supposed to be the escape from the HBM bottleneck. The CS-6 design concedes that the escape still runs through DRAM, and the plant SK Hynix just broke ground on is where a lot of it will come from. SourcesBA
Three-quarters of new dramas on China's TikTok are now AI-made
AI-generated short dramas have gone from a novelty to the majority of new content on Douyin: in August 2025 there were 655 new AI dramas in a month, and a year later that number was 74,313, about 75% of all new titles, at a pace of roughly one every 36 seconds. This is the first mass consumer application of AI-generated video, running at industrial scale in a $16.5 billion micro-drama market, and it is already producing industrial problems: platforms are moving to standardize AI character creation to stop everything looking the same, and industry estimates suggest only about one in 77 of these titles breaks even. It is the clearest picture available of what happens when the cost of producing video falls toward zero. The answer is not better shows; it is a glut, a platform scrambling to impose quality rules, and an economics where almost nothing recoups. Every media industry watching generative video should read the Chinese short-drama feed as the near future of their own supply. SourcesBB
A world-model startup tripled to $6 billion in eight weeks
General Intuition, a New York startup that trains world models on hundreds of millions of hours of video-game footage, is raising at a $6 billion , nearly triple the $2.3 billion it was worth eight weeks ago, with Valor Equity Partners, Point72 Ventures and Seven Seven Six joining existing backers Khosla and General Catalyst. The company spun out of the gameplay-clip platform Medal in October 2025 and is pushing from game agents into robotics, planning to train physical embodiments on CoreWeave compute while growing from 12 employees to around 25 by year-end. The number to be skeptical of is the velocity: a 2.6x markup in two months, on a company with a dozen people and no shipped robot, is the market pricing the world-model thesis, not the traction. But the thesis is real, that gameplay footage is a cheap, abundant source of the spatial and temporal reasoning robots need, and the round is a bet that whoever learns physics from games gets to embodiment faster than whoever waits for real-world demonstrations. SourcesB